clawbench-cli 0.1.2__py3-none-any.whl

clawbench/data/test-cases/785-beauty-personal-care-skincare-curology/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 785,
+    "metaclass": "beauty-personal-care",
+    "class": "skincare",
+    "description": "On Curology, complete the skin condition questionnaire (skin type, acne, aging concerns) and reach the checkout or subscription page with payment details filled in.",
+    "sites_involved": [
+      "curology.com"
+    ],
+    "platform": "curology",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Curology, complete the skin condition questionnaire (skin type, acne, aging concerns) and reach the checkout or subscription page with payment details filled in.",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/788-beauty-personal-care-makeup-the-ordinary/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 788,
+    "metaclass": "beauty-personal-care",
+    "class": "makeup",
+    "description": "Complete the skin type quiz on The Ordinary and add the two recommended products to cart",
+    "sites_involved": [
+      "theordinary.com"
+    ],
+    "platform": "the-ordinary",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "Complete the skin type quiz on The Ordinary and add the two recommended products to cart",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/789-beauty-personal-care-makeup-fenty-beauty/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 789,
+    "metaclass": "beauty-personal-care",
+    "class": "makeup",
+    "description": "On Fenty Beauty's website, add Pro Filt'r Soft Matte Foundation (shade 185N) and Invisible Setting Powder to cart and begin checkout",
+    "sites_involved": [
+      "fentybeauty.com"
+    ],
+    "platform": "fenty-beauty",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Fenty Beauty's website, add Pro Filt'r Soft Matte Foundation (shade 185N) and Invisible Setting Powder to cart and begin checkout",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/793-beauty-personal-care-beauty-retail-mac-cosmetics/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 793,
+    "metaclass": "beauty-personal-care",
+    "class": "beauty-retail",
+    "description": "On MAC Cosmetics, book an in-store makeup service at a California Los Angeles location and proceed to the payment page",
+    "sites_involved": [
+      "maccosmetics.com"
+    ],
+    "platform": "mac-cosmetics",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On MAC Cosmetics, book an in-store makeup service at a California Los Angeles location and proceed to the payment page",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/794-beauty-personal-care-salon-booking-styleseat/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 794,
+    "metaclass": "beauty-personal-care",
+    "class": "salon-booking",
+    "description": "On StyleSeat, search for hairstylists in your city, select one rated 4.8 stars or above, and book a haircut appointment at the earliest available time slot",
+    "sites_involved": [
+      "styleseat.com"
+    ],
+    "platform": "styleseat",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On StyleSeat, search for hairstylists in your city, select one rated 4.8 stars or above, and book a haircut appointment at the earliest available time slot",
+  "eval_schema": {
+    "url_pattern": "styleseat\\.com/api/v1/appointment-request/",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/795-pet-animal-care-pet-adoption-aspca/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 795,
+    "metaclass": "pet-animal-care",
+    "class": "pet-adoption",
+    "description": "On the ASPCA adoption page, filter adoptable dogs under 2 years old, select one, and submit a complete adoption application with personal info and living situation",
+    "sites_involved": [
+      "aspca.org/adopt-pet"
+    ],
+    "platform": "aspca",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On the ASPCA adoption page, filter adoptable dogs under 2 years old, select one, and submit a complete adoption application with personal info and living situation",
+  "eval_schema": {
+    "url_pattern": "global\\.api\\.prd\\.adopets\\.app/adopter/register",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/796-pet-animal-care-pet-supplies-grooming-petsmart/extra_info/pet_info.json

@@ -0,0 +1,12 @@
+{
+  "name": "Maple",
+  "species": "Dog",
+  "breed": "Golden Retriever",
+  "age_years": 4,
+  "weight_lbs": 65,
+  "sex": "Female (spayed)",
+  "date_of_birth": "2022-03-15",
+  "vaccinations_up_to_date": true,
+  "microchip_id": "985121012345678",
+  "dietary_notes": "Grain-free kibble, sensitive stomach"
+}

clawbench/data/test-cases/796-pet-animal-care-pet-supplies-grooming-petsmart/task.json

@@ -0,0 +1,30 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 796,
+    "metaclass": "pet-animal-care",
+    "class": "pet-supplies-grooming",
+    "description": "On PetSmart, order dog food for same-day delivery and book a professional grooming appointment for the upcoming Saturday",
+    "sites_involved": [
+      "petsmart.com"
+    ],
+    "platform": "petsmart",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On PetSmart, order dog food for same-day delivery and book a professional grooming appointment for the upcoming Saturday",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": [
+    {
+      "path": "extra_info/pet_info.json",
+      "description": "Pet information for Alex Green's dog Maple (Golden Retriever, 4 years, 65 lbs)"
+    }
+  ]
+}

clawbench/data/test-cases/799-pet-animal-care-pet-insurance-aspca-pet-health-insurance/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 799,
+    "metaclass": "pet-animal-care",
+    "class": "pet-insurance",
+    "description": "On ASPCA Pet Health Insurance, get a quote for a 2-year-old cat by entering breed, age, and ZIP code, then select a plan and enroll",
+    "sites_involved": [
+      "aspcapetinsurance.com"
+    ],
+    "platform": "aspca-pet-health-insurance",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On ASPCA Pet Health Insurance, get a quote for a 2-year-old cat by entering breed, age, and ZIP code, then select a plan and enroll",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/801-pet-animal-care-pet-friendly-travel-bringfido/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 801,
+    "metaclass": "pet-animal-care",
+    "class": "pet-friendly-travel",
+    "description": "On BringFido, search for pet-friendly hotels in San Francisco that allow large dogs, select one, and complete a booking reservation",
+    "sites_involved": [
+      "bringfido.com"
+    ],
+    "platform": "bringfido",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On BringFido, search for pet-friendly hotels in San Francisco that allow large dogs, select one, and complete a booking reservation",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/803-pet-animal-care-pet-medical-pawp/extra_info/pet_info.json

@@ -0,0 +1,12 @@
+{
+  "name": "Maple",
+  "species": "Dog",
+  "breed": "Golden Retriever",
+  "age_years": 4,
+  "weight_lbs": 65,
+  "sex": "Female (spayed)",
+  "date_of_birth": "2022-03-15",
+  "vaccinations_up_to_date": true,
+  "microchip_id": "985121012345678",
+  "dietary_notes": "Grain-free kibble, sensitive stomach"
+}

clawbench/data/test-cases/803-pet-animal-care-pet-medical-pawp/task.json

@@ -0,0 +1,30 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 803,
+    "metaclass": "pet-animal-care",
+    "class": "pet-medical",
+    "description": "On Pawp, sign up for a membership and schedule an online vet consultation for your pet, describe the symptoms, and submit the consultation request",
+    "sites_involved": [
+      "pawp.com"
+    ],
+    "platform": "pawp",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Pawp, sign up for a membership and schedule an online vet consultation for your pet, describe the symptoms, and submit the consultation request",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": [
+    {
+      "path": "extra_info/pet_info.json",
+      "description": "Pet information for Alex Green's dog Maple (Golden Retriever, 4 years, 65 lbs)"
+    }
+  ]
+}

clawbench/data/test-cases/807-pet-animal-care-pet-dna-embark/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 807,
+    "metaclass": "pet-animal-care",
+    "class": "pet-dna",
+    "description": "On Embark's website, add the Breed + Health Kit to cart, proceed to checkout, and fill in shipping information without submitting payment",
+    "sites_involved": [
+      "embarkvet.com"
+    ],
+    "platform": "embark",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Embark's website, add the Breed + Health Kit to cart, proceed to checkout, and fill in shipping information without submitting payment",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/809-pet-animal-care-pet-adopt-petfinder/task.json

@@ -0,0 +1,28 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 809,
+    "metaclass": "pet-animal-care",
+    "class": "pet-adopt",
+    "description": "Search for adoptable medium-sized dogs on Petfinder and submit an adoption inquiry form",
+    "sites_involved": [
+      "petfinder.com"
+    ],
+    "platform": "petfinder",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "Search for adoptable medium-sized dogs on Petfinder and submit an adoption inquiry form",
+  "eval_schema": {
+    "url_pattern": "psl\\.petfinder\\.com/graphql",
+    "method": "POST",
+    "body": {
+      "operationName": "CreateAdoptionInquiry"
+    }
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/812-pet-animal-care-pet-subscription-ollie/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 812,
+    "metaclass": "pet-animal-care",
+    "class": "pet-subscription",
+    "description": "Complete dog profile on Ollie and subscribe to a custom fresh food meal plan",
+    "sites_involved": [
+      "myollie.com"
+    ],
+    "platform": "ollie",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "Complete dog profile on Ollie and subscribe to a custom fresh food meal plan",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/815-personal-management-records-mgmt-myheritage/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 815,
+    "metaclass": "personal-management",
+    "class": "records-mgmt",
+    "description": "Register on MyHeritage and order a DNA test kit, enter shipping address (123 Main St, Toronto, ON M5V 1A1), and complete the order submission",
+    "sites_involved": [
+      "myheritage.com"
+    ],
+    "platform": "myheritage",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "Register on MyHeritage and order a DNA test kit, enter shipping address (123 Main St, Toronto, ON M5V 1A1), and complete the order submission",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/821-education-learning-reading-self-study-blinkist/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 821,
+    "metaclass": "education-learning",
+    "class": "reading-self-study",
+    "description": "Register on Blinkist with email reader2026@example.com, subscribe to the Premium annual plan, then search for Atomic Habits and add it to My Library (favorites)",
+    "sites_involved": [
+      "blinkist.com"
+    ],
+    "platform": "blinkist",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "Register on Blinkist with email reader2026@example.com, subscribe to the Premium annual plan, then search for Atomic Habits and add it to My Library (favorites)",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/861-entertainment-hobbies-movies-cineplex/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 861,
+    "metaclass": "entertainment-hobbies",
+    "class": "movies",
+    "description": "On Cineplex, purchase 2 adult evening tickets for any currently showing action or sci-fi movie at Scotiabank Theatre Toronto (259 Richmond St W), choose a showtime after 7 PM on the upcoming Saturday, and select two adjacent seats in row G or H in the center section",
+    "sites_involved": [
+      "cineplex.com"
+    ],
+    "platform": "cineplex",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Cineplex, purchase 2 adult evening tickets for any currently showing action or sci-fi movie at Scotiabank Theatre Toronto (259 Richmond St W), choose a showtime after 7 PM on the upcoming Saturday, and select two adjacent seats in row G or H in the center section",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/862-entertainment-hobbies-movies-amc-theatres/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 862,
+    "metaclass": "entertainment-hobbies",
+    "class": "movies",
+    "description": "On AMC Theatres, buy 2 IMAX tickets for the highest-rated movie currently showing at AMC Empire 25 in New York, select an upcoming Friday evening showtime, and choose two adjacent center seats in rows E–H",
+    "sites_involved": [
+      "amctheatres.com"
+    ],
+    "platform": "amc-theatres",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On AMC Theatres, buy 2 IMAX tickets for the highest-rated movie currently showing at AMC Empire 25 in New York, select an upcoming Friday evening showtime, and choose two adjacent center seats in rows E–H",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/864-entertainment-hobbies-show-tickets-ticketmaster/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 864,
+    "metaclass": "entertainment-hobbies",
+    "class": "show-tickets",
+    "description": "On Ticketmaster Canada, purchase 2 tickets to the next available concert at Scotiabank Arena in Toronto, select seats in the lower bowl (sections 100–120), preferring center-stage view",
+    "sites_involved": [
+      "ticketmaster.ca"
+    ],
+    "platform": "ticketmaster",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Ticketmaster Canada, purchase 2 tickets to the next available concert at Scotiabank Arena in Toronto, select seats in the lower bowl (sections 100–120), preferring center-stage view",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/865-travel-outdoor-hipcamp/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 865,
+    "metaclass": "travel",
+    "class": "outdoor",
+    "description": "On Hipcamp, book a 2-night tent campsite near Lake Tahoe for the upcoming weekend (Friday–Sunday), filter for sites with lake access and campfire allowed, select the highest-rated available site",
+    "sites_involved": [
+      "hipcamp.com"
+    ],
+    "platform": "hipcamp",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Hipcamp, book a 2-night tent campsite near Lake Tahoe for the upcoming weekend (Friday–Sunday), filter for sites with lake access and campfire allowed, select the highest-rated available site",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/867-entertainment-hobbies-movies-fandango/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 867,
+    "metaclass": "entertainment-hobbies",
+    "class": "movies",
+    "description": "On Fandango, purchase 3 tickets for any currently showing thriller at Regal Union Square in New York, select an upcoming Friday evening showtime after 6 PM, choose 3 adjacent seats in rows D–F center section",
+    "sites_involved": [
+      "fandango.com"
+    ],
+    "platform": "fandango",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Fandango, purchase 3 tickets for any currently showing thriller at Regal Union Square in New York, select an upcoming Friday evening showtime after 6 PM, choose 3 adjacent seats in rows D–F center section",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/872-daily-life-food-opentable/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 872,
+    "metaclass": "daily-life",
+    "class": "food",
+    "description": "On OpenTable, book a dinner reservation at a Japanese omakase restaurant in San Francisco for 2 people, the upcoming Thursday at 8 PM, with counter/bar seating preference, rating 4.5+",
+    "sites_involved": [
+      "opentable.com"
+    ],
+    "platform": "opentable",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On OpenTable, book a dinner reservation at a Japanese omakase restaurant in San Francisco for 2 people, the upcoming Thursday at 8 PM, with counter/bar seating preference, rating 4.5+",
+  "eval_schema": {
+    "url_pattern": "opentable\\.com/dapi/fe/gql\\?optype=mutation&opname=BookDetailsExperienceSlotLock",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/873-daily-life-food-resy/task.json

@@ -0,0 +1,28 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 873,
+    "metaclass": "daily-life",
+    "class": "food",
+    "description": "On Resy, find and book a Michelin-recommended Italian restaurant in Manhattan for 2 people, the upcoming Friday at 8:30 PM, specify dining room (not bar) seating",
+    "sites_involved": [
+      "resy.com"
+    ],
+    "platform": "resy",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Resy, find and book a Michelin-recommended Italian restaurant in Manhattan for 2 people, the upcoming Friday at 8:30 PM, specify dining room (not bar) seating",
+  "eval_schema": {
+    "url_pattern": "api\\.resy\\.com/3/details",
+    "method": "POST",
+    "body": {
+      "commit": 1
+    }
+  },
+  "time_limit": 30,
+  "extra_info": []
+}

clawbench/data/test-cases/876-entertainment-hobbies-show-tickets-vivid-seats/task.json

@@ -0,0 +1,25 @@
+{
+  "$schema": "../task.schema.json",
+  "metadata": {
+    "task_id": 876,
+    "metaclass": "entertainment-hobbies",
+    "class": "show-tickets",
+    "description": "On Vivid Seats, purchase 2 orchestra center tickets for Hamilton on Broadway at the Richard Rodgers Theatre, select the earliest available performance date",
+    "sites_involved": [
+      "vividseats.com"
+    ],
+    "platform": "vivid-seats",
+    "common_info": {
+      "email_credentials": "credentials to use the assigned disposable email account",
+      "user_info": "alex_green_personal_info.json; the dummy user's personal information",
+      "user_resume": "PDF resume with disposable email account injected"
+    }
+  },
+  "instruction": "On Vivid Seats, purchase 2 orchestra center tickets for Hamilton on Broadway at the Richard Rodgers Theatre, select the earliest available performance date",
+  "eval_schema": {
+    "url_pattern": "__PLACEHOLDER_WILL_NOT_MATCH__",
+    "method": "POST"
+  },
+  "time_limit": 30,
+  "extra_info": []
+}